Any thoughts on how far into the future to forecast? I'm working on forecasting our web monthly traffic in R. I'm using Holt-Winters and the dates Jan 2015 - Sep 2017. I want to forecast the next 12 months. I know the further I get out, the less accurate I am. Is there a good rule/ratio to follow? Thanks!
Add a comment
|
3 Answers
$\begingroup$
$\endgroup$
0
Probably your forecast should include both a point forecast and a variability/confidence band around the point. The variability band generally should be increasing as you forecast farther out into the future. If your forecast incorporates this band there should be a point beyond which your forecast becomes not useful because the variability becomes too high.
answered Oct 11, 2017 at 14:28
$\begingroup$
$\endgroup$
0
Long term forecast for web traffic is really quite difficult with full generality and I don't know of efficient method for long term prediction (several months). In many cases Holt-Winters can't even extract useful yearly cycles because you need several years, and anyway the world, trends and fashions change from one year to the next. Holt-Winters works very well on weekly cycles but not further as far as I have observed.
Also, it looks like every website or webpage behave quite differently, there is a lot of noise and a lot of impredictability.
Basically, the error will increase as you go forward, a bit like a basic random walk prediction (forecasted as $\hat y_{t+h}=y_t$).
If you have many time-series, you can calculate an average error for the method that you use, using predicted vs real (with a burn-in safety to ignore HW initializing). However there is no reason that the observations of errors on the past of a single time-series will generalize to the future.
Maybe you can compare it to the standard error curve to the one obtained by "silly" forecast $\hat y_{t+h}=y_t$ and see how better you are.
$\begingroup$
$\endgroup$
0
Given that your coefficients are statistically significant and your model's errors have passed multiple "model specification tests" then I would obtain confidence limits by re-samplimg the model's errors via Monte Carlo while also incorporating the psi weights to possibly inflate the forecast variance due to auto-projection.
The width of the limits will suggest to you "how far ahead you can safely forecast"
